Data Driven Approach To Protect Public Figures Against Deepfakes
Raksha Varahamurthy · Deep Blue (University of Michigan) · 2024
The proliferation of deepfake technology poses a significant threat to public figures, particularly political leaders. This thesis presents a data-driven approach to protect public figures against deepfakes, addressing critical gaps in open-source and commercial datasets by compiling diverse audio-visual data from platforms such as YouTube and Audible. Custom models are trained on these datasets, focusing on speaker-specific nuances and emotional recognition. This research advances deepfake detection by capturing hidden emotions, such as happiness, sadness, anger, and laughter, and conducts a comprehensive analysis of deepfake generation platforms. The study proposes regulatory measures to mitigate deepfake misuse, offering practical solutions grounded in ethical considerations. The findings contribute to the state-of-the-art in deepfake research and lay the groundwork for future advancements in protecting public figures from deepfake-related harm.